Abstract
This paper presents a problem of model learning to navigate a ball to a goalstate in a circular maze environment with two degrees of freedom. Motion of theball in the maze environment is influenced by several non-linear effects suchas friction and contacts, which are difficult to model. We propose a hybridmodel to estimate the dynamics of the ball in the maze based on GaussianProcess Regression equipped with basis functions obtained from physic firstprinciples. The accuracy of the hybrid model is compared with standardalgorithms for model learning to highlight its efficacy. The learned model isthen used to design trajectories for the ball using a trajectory optimizationalgorithm. We also hope that the system presented in the paper can be used as abenchmark problem for reinforcement and robot learning for its interesting andchallenging dynamics and its ease of reproducibility.